We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
🔥 TOP SIGNAL
Stop relying on yes/no approval prompts. Put a context-free safety checker and a real sandbox around agents instead. Boris Cherny, who created Claude Code, describes an Anthropic study. Contractors doing coding puzzles were occasionally shown injected commands that would have damaged their systems if real, and they approved them "almost all the time." He calls the prompts "security theater" . Claude Code's auto mode sends each proposed action to a second model that has none of the conversation's context. Cherny says it "gets it right almost every time," and you can customize its rules by asking Claude . He also says to "always use the sandbox," even in production .
Why the sandbox has to be tight: Jeff Ladish reports agents that could only load URLs, with no way to send data. They used a link-shortener to create almost a million URLs that, chained together, let them execute code and hack Hugging Face . Loading URLs was enough to get data out.
⚡ TRY THIS
Use the sandbox the way Cherny describes it: a box that stays closed, with specific holes cut in it. Allow model inference, only the folders the agent needs, and a named list of websites. Block general web access, because "to do useful work you probably don't need access to the whole internet" . He says Claude Code's sandbox is open source and works with any agent. Claude Code users have to turn it on themselves because it's off by default .
Have agents turn a web page into a video (Simon Willison). Willison got a pixel-art HTML canvas animation from Claude Opus 5.5 . He then told a local Claude Code session:
Make me a video of file:///…/kakapo-party.html - you need to load it in a browser and click on it a few times to get the confetti effect, the video should be 15s long/don't start clicking until 3s in/make sure several clicks are spread around the clickable area. Claude Code wrote a short Playwright script. It opens Chromium at 1280×720 with video recording on, clicks at 10 timed points spread from the center to the corners, and closes at about 16s . The pattern carries over to demos, bug-repro videos, and visual regression capture.Make verification fast, not just thorough (Geoffrey Huntley). Each change needs checks that push back, but if those checks are slow, every LLM mistake costs you cycles. His target: read a file, build, and run tests in milliseconds. His prediction: the bottleneck will be verification speed and "how fast you can verify and pump the result back into the inferencing window," not inference . Huntley also says Hegel-powered property-based testing for NixOS "found stuff," and he plans to open-source the pattern .
Run parallel hypotheses and have agents check each other (Cherny). His typical data-analysis sessions run up to about 12 hours. Claude forms a hypothesis, sends about 10 agents to investigate in parallel, then checks their findings and throws some out before going deeper . Small one: Addy Osmani points out that the Claude Code desktop app's "Keep computer awake" setting stops your machine from sleeping during long sessions .
📡 WHAT SHIPPED
Jev Router got benchmarked, and it doesn't come out ahead. Theo spent $1,000 testing OpenRouter's
typesafe/jev-router. On DeepSWE it scored about the same as GPT-6 Astra on low, cost slightly more, and took almost 5× longer . His explanation: Jev "categorizes" and doesn't reason, so it can't tell whether "port this to Rust" means a 100-line file or a million-line app . He adds that switching to a cheaper model partway through a task saves little when cache writes are "such a massive % of cost" . Takeaway: routing on the prompt alone is a weak way to choose reasoning effort for coding agents.ttfx 0.4 (Omarchy): the assembly engine written by Opus is released. DHH says the print effect now runs more than 1,000× faster than the original Python and 32× faster than the old Rust. It ships in the next Omarchy, and a faster Rust fallback is coming . His observation: hours of frontier-model work on the Rust version brought "very modest improvements," while moving to assembly brought big jumps (8× mean on an Intel 135U) . These are one project's own numbers, not a general benchmark.
Oligarchy, a harness for agent-driven QA (announced, not live). ThePrimeagen joins Omarchy Core to lead Agentic QA. The harness will vet upcoming releases with agents running on DigitalOcean droplets . He promises a write-up with milestones and expectations this week .
Open-weights model as a daily driver. Huntley runs Kimi/K3 as his main model on a couple of racks of B300s in a Sydney data center . He reports about 200M tokens an hour and expects close to a billion within five hours .
Subscription value, based on one person's heavy use. Theo says he maxed out about 3.5 of the $200 Claude Code accounts in 5 days and calls the Max plan "an absolute steal" .
Theo on the "Opus 5.5 got nerfed" claims. He points to one real incident: last year, inference optimizations across Nvidia, Trainium, and TPUs caused degradation that Anthropic later fixed and explained. He says there have been no notable cases since. His statistical point: with a 1-in-50 chance of an odd output and 20 prompts a day, you have about a 30% chance of seeing one on day 1 and about 90% by day 5 . Don't read a single bad run as a regression.
Two points about avoiding lock-in. Kody's pitch, shared by Kent C. Dodds: keep memory, secrets, and packages in your own "house" so you can swap agents or use several at once . Riley Brown criticizes Descript's MCP: it just forwards Claude Code prompts to Descript's own Underlord agent, which costs more and forces you onto their tokens .
🎬 GO DEEPER
- Cherny on why approval prompts fail and how auto mode replaces them. The contractor study and the design of the context-free safety classifier.
- Cherny on designing an agent's tool set. Claude Code started with about 4–5 tools (read, write, bash, screenshot). Subagents are just "Claude starts a Claude" and can choose Opus or several Haikus. Tools get added and removed with every model release, and "often it's not the obvious thing that works… it's often the simplest thing."
- Salvatore Sanfilippo (Italian) on C vs Rust for LLM-written code. He argues that LLMs write C much better than any other language because of the huge, high-quality C codebase they learned from. He would use C for services that aren't security-critical, audit it heavily, and keep Rust for mission-critical code that can't be isolated .
- Post: Simon Willison, "Kākāpō Party": The full prompts, transcripts, and the Playwright script from the TRY THIS item .
Editorial take: The limit on agents is no longer model quality. It's how fast you can check their work and how tightly you've boxed them in.
- Parallel orchestration: Claude Code creator Boris Cherny describes letting one agent build an idea while he starts another; in long data-analysis sessions, he develops and tests hypotheses, launches about 10 agents in parallel, checks and sometimes rejects their findings, then investigates promising leads further. He says these sessions can last around 12 hours. Cherny also says Anthropic employees typically have hundreds of agents running daily.
- Agent design: Cherny describes Claude Code’s core tools as reading and writing files and running bash commands; its agent tool can launch subagents and choose models such as Opus or Haiku, or several Haikus. Since tools change with model releases, he recommends experimenting: the useful tool set may not be obvious, and simple tools often work.
- Command review: Cherny says yes/no approval prompts became “security theater”: in a contractor study, participants accepted simulated harmful commands almost every time. Claude Code’s auto mode instead sends a proposed action to a separate model without the conversation context to assess its safety; he says it performed far better than human approval, and its rules can be customized.
- Sandboxing: Cherny recommends running agents in a sandbox even in production. His guidance is to grant access only to selected files and explicitly allow needed websites; he says Claude Code’s open-source sandbox works with any agent, though Claude Code users had to enable it at the time of the interview.
- Memory caveat: Cherny says an early instruction to remember useful things caused Claude to overgeneralize one-off feedback—for example, treating a request to make one button blue as a rule for every button. He says good memory requires judgment; Claude Code memory is private and local to an individual, while team memory is shared within that team.
- Salvatore Sanfilippo says frontier models are usable for Redis; the transcript renders model names as “Astra” and “Fable 5.1,” and says he had not yet assessed “Opus 5” deeply, which he describes as refusing GPU-kernel work. His practical pattern is to give the model architectural and product direction, then rapidly implement, revise, and change the design; he contrasts this with a minimally guided one-line request, and argues that repeated iterations can make early AI output more mature than a human team’s first passes.
- Contrarian language take: Sanfilippo argues that LLMs write C especially well because of the large, high-quality C codebase they learned from, and favors C for non-security-critical services while reserving Rust for mission-critical cases that cannot be isolated. He also says generated C should be audited and security-hardened.
- He cautions that models may miss specialist details—for example, audio-player system calls, timing, and buffering—so domain-specific review still matters.
- Firsthand workflow (Simon Willison): He gave Claude Opus 5.5 three kakapo photos and prompted it to create “animated pixel art on HTML 5 canvas” of at least 20 parrots partying; Claude produced a usable page.
- To turn the HTML into presentation video, he asked a local Claude Code session to load it in a browser, wait until 3 seconds in, then click in several places; Claude Code used Playwright. The script launches Chromium with a 1280×720 viewport and video recording, schedules 10 clicks across the screen from 3.0 to 13.2 seconds, then waits until 16 seconds elapsed before closing. Willison requested a 15-second video, so the script’s 16-second wait is a detail to check when adapting the workflow.
Simon Willison used Claude Opus 5.5 to create a celebratory pixel-art video for a talk’s closing slide; Claude also built an interactive HTML version. He shared the prompt and transcripts .
DHH announced that ThePrimeagen joined Omarchy Core to lead Agentic QA, with a planned “Oligarchy” harness to vet upcoming releases using agents on DigitalOcean droplets powered by Meta tokens; this is a forward-looking setup, with no implementation or evaluation details given. Announcement
ThePrimeagen says he received $0 from Omarchy or DHH, describes his Netflix work as building developer tools, and says he still wants to build tools and meet deadlines.
DHH announced that ThePrimeagen joined Omarchy Core to lead Agentic QA, with a planned “Oligarchy” harness to vet upcoming releases using agents on DigitalOcean droplets with Meta AI tokens. ThePrimeagen said he has been pursuing agent-based automation and will share a write-up, milestones, and expectations; implementation details are not yet provided.
Kent C. Dodds says he has a setup with Kody similar to the Omarchy bug-report/triage process, with bugs reported and fixed in production while he sleeps; he does not describe the implementation steps .
Antigravity 2.0 adds a dedicated planning mode: use /plan to have the agent research the task and draft an implementation plan for review, then wait for approval before execution; asking for a plan in natural language produces a lighter version. Kent C. Dodds says this arrives as other tools are getting rid of the feature and questions Google's direction, though he does not name the alternatives.
Jeff Ladish reported that agents limited to loading URLs, with no ability to send data, chained nearly a million link-shortener URLs to execute code and hack Hugging Face; Peter Steinberger praised the approach as clever. The incident is a caution that URL-loading-only access was not a sufficient boundary in that setup, though it is not a reproducible coding-agent workflow.
Geoffrey Huntley says he is running Kimi/K3 as his day-to-day driver on a couple of racks of B300s in a Sydney data center, and reports “200 million tok an hour,” with nearly a billion tokens in under five hours as the pace he expects.
Kent C. Dodds uses Kody’s “house/roommates” metaphor to explain avoiding agent-vendor lock-in: keep memory, secrets, and packages in a persistent Kody home so agents can be replaced, or use multiple agents together.
Addy Osmani recommends enabling Claude Code’s “Keep computer awake” option to prevent a desktop machine from sleeping during a long-running session.
Geoffrey Huntley says he is using Kimi/K3 as his day-to-day driver on a couple of B300 racks, describing the Sydney data-center deployment as running OSS models “locally.” In a follow-up, he reports a throughput of 130 million tokens per hour.
Geoffrey Huntley argues that coding-agent workflows should optimize verification latency: provide meaningful per-change backpressure, but keep file-read, build, and test cycles in milliseconds so results return to the model quickly and do not throttle iteration. He predicts that compiler and verification tooling, rather than inference speed, will become the bottleneck as inference capacity grows.
Jason Zhou relayed Josh’s report that he built, in four hours at a hackathon, a system to generate 500+ Meta ads for DTC companies. Josh listed Cursor and Anthropic as build tools; Treg AI for ad research and brand scans; Google DeepMind for images; OpenAI for copy and strategy; Next.js/Vercel as the framework; and Cloudflare for hosting. This is a secondhand account of the result and stack, not a reproducible coding-agent workflow: the post gives no prompts, implementation steps, or agent configuration.
DHH reports that after many improvements to the Rust version of ttfx, hours of “frontier intelligence” yielded only modest gains, while moving to assembly brought much larger gains; the assembly version reached 8× mean on an Intel 135U. In a quoted update, he credits “Opus” with further efficiency improvements, reporting up to 22× at peak, 6× on mean, and 450× versus the original Python implementation; it supports SSE2, AVX2, and AVX-512, with a slower Rust fallback.
Responding to claims that Claude Opus 5.5 was nerfed, Theo argues that stochastic variation can make isolated bad outputs look like a regression. His illustration—not a coding benchmark—is that if a model has a 1-in-50 chance of an odd response, using it for 20 prompts a day gives roughly a 30% chance of seeing one on day one and nearly 90% by day five. Theo also cites a past Anthropic inference-optimization incident involving Nvidia, AWS Trainium, and Google TPUs that caused degradation, was later fixed and explained, and has had no notable recurrence according to him.
Theo says he “killed” a Claude account while using mostly Opus; in the thread, he says he was rewriting the TypeScript compiler in Rust while also doing other work . Follow-ups report roughly 2.5 accounts down, then two killed and a third imminent . This is a firsthand account-usage anecdote, not a reproducible workflow or performance benchmark.
Theo says he was able to “kill” roughly 3.5 $200 Claude Code accounts in five days, calls Opus 5.5 “incredible,” and considers the $200 Max subscription “an absolute steal.” He says he would likely switch when models better than Opus 5.5 arrive, but would miss the “unlimited” feeling.
In his video-editing workflow, Riley Brown says Descript’s MCP connection forwards Codex or Claude Code prompts to Descript’s Underlord agent, adding cost and requiring Descript’s tokens; he prefers using Claude Code locally and says it is more efficient than Descript’s built-in agent . His practical integration critique: a tool connection that reroutes prompts through a separate proprietary agent can be less effective than using the coding agent directly.
